Impact of the number of dates and their sampling on a NDVI time series reconstruction methodology to monitor urban trees with Ven[formula omitted]s satellite

Abstract

This article studies the influence of the number of satellite remote sensing acquisition dates and their sampling on the performance of a time series reconstruction method developed in Granero-Belinchon et al. 2020. This method initially aimed at monitoring urban London plane (Platanus x acerifolia) trees, and was tested with Sentinel-2 imagery at spatial resolutions of 10 and 20 m and a temporal revisit of 5 days. Due to its higher revisit frequency of 2 days while having a similar spatial resolution of 10 m, Venμs imagery was consequently used in the present article to fulfill with the purpose of this study. The strategy relies on the building of different acquisition date configurations based on the Venμs time series by considering uniform and non-uniform samplings and with a total number of acquisitions ranging from 45 to 14. Thus, the aim of the article is to examine the number of annual acquisitions needed to describe properly a vegetation phenological cycle and the impact of the annual sampling of these acquisitions on the final reconstructed time series. To this end, this study was carried out by using the widely used Normalized Difference Vegetation Index (NDVI). Results showed that on one hand, applied on an acquisition configuration composed of at least 18 uniformly sampled dates throughout the year, this reconstruction methodology is able to describe correctly the annual NDVI dynamics but leads to inaccuracies in the description of intra-annual ones. Nevertheless, these intra-annual descriptions are improved with the increase of the number of acquisitions. On the other hand, strongly non-uniform acquisition date samplings lead to inaccurate descriptions of the undersampled time periods but correct descriptions of the rest of the time series curve. The study case is London planes located in Toulouse (France) with 45 cloud-free Venμs images during the year 2019. Finally, this work emphasizes the main limitations of the studied reconstruction methodology when few acquisitions or very non-uniform acquisition date samplings are available and thus the identification of borderline cases in future applications and other study cases.